Genesis Built a 250,000-Line C Compiler for $44

Genesis, an agent orchestration system, built a working C compiler in Rust across more than 120 hours for $44 in model tokens, according to a paper posted to arXiv. The result is about 250,000 tracked lines of code that pass the complete c-testsuite plus most of the LLVM and Csmith test suites. It started from a repo with no compiler in it at all.

The system's move is keeping the project alive longer than any single agent working on it. Each agent operates inside a local world anchored to an accepted code version and a specific path in the repo. It proposes a change, and that change only becomes real once it's accepted into the shared version history. Work delegates recursively across paths, so the 1,000-plus agent episodes behind the compiler never had to hold all 250k lines in context at once. DeepSeek V4 Flash was the model behind all of it.

A second build, a real speedup

Genesis also reimplemented 13 modules of MESA, a stellar astrophysics code, turning more than 100,000 lines of Fortran into a roughly 90,000-line Rust workspace. Across six numerical workloads, median speedups ranged from 1.55 to 6.87 times the original, per the paper.

The paper doesn't say how many of those 1,000-plus episodes got rejected before anything landed in the accepted history. There's no way to tell from this account how much wasted agent time sits behind the $44 figure.

The number worth arguing about isn't $44. It's that zero lines of agent memory survive between episodes, so the version history has to do all the remembering. A related 5,280-episode study found that guards checking only an agent's local state missed authority-laundering attempts in 22 of 96 episodes. Guards checking version-history provenance caught more of them. Genesis leans on the same idea: trust the accepted history, not the agent.